At the Computex trade show in Taipei, major GPU makers Nvidia, Intel, and AMD shared details about new desktop GPUs that let users run generative AI models locally. Nvidia announced that its DGX Spark desktop (formerly Project DIGITS), a small box that sits on the desktop, will ship in July from top PC makers including Dell, HP, Acer, Asus, and Lenovo. The system includes Nvidia’s latest Blackwell GPU and delivers 1 petaflops of performance. (See “Nvidia’s Project DIGITS puts AI supercomputing on the desktop” for details.) It is “your own AI cloud sitting right next to you, and it’s always on, always waiting for you,” said Jensen Huang, Nvidia’s CEO, during a keynote at the Computex trade show. While Spark is targeted exclusively at AI, Intel and AMD announced GPUs that mix in AI and graphics. AI typically requires special circuitry on GPUs capable of matrix multiplications. AMD’s Radeon 9700 AI Pro GPU has 128 AI accelerators and can run a DeepSeek R1 reasoning model with up to 32 billion parameters. Desktops configured with four Radeon 9700 GPUs can run Mistral’s Large Instruct model with 123 billion parameters. The GPU will be available in July. Intel’s Arc Pro B50, which is targeted at workstations and priced at $299, is a budget-friendly GPU aimed primarily at graphics, but also has 128 specialized AI engines. The GPU can run DeepSeek R1 with up to 14 billion parameters. It doesn’t have the memory to run a model with 32 billion or more parameters, according to a slide deck provided by Intel. The Arc Pro B50 has 16GB of memory — half that of AMD’s Radeon 9700 AI Pro. The Intel GPU draws 70 watts of power, while AMD’s GPU draws 300 watts. AI development is moving toward personal computing and away from expensive time-share cloud computing resources, said Jack Gold, principal analyst at J. Gold Associates. “They may still run the output in the cloud at production, but increasingly development is moving towards PCs, especially as we see AI-powered development tools being used to develop AI programs,” Gold said. Chip makers also see AI workstations — which are primarily sold for graphics and gaming — as the next big thing in high-end PC sales. “Discrete GPUs in desktops and laptops is a good market, but the potential for AI PCs is probably much higher in numbers. It’s a good expansion for the PC market suppliers,” Gold said. AI laptops such as Windows-based Copilot+ PCs have neural processing units, which are for AI consumption and not development, Gold noted. AI is increasingly becoming a secondary workload in workstations, especially if they are running Nvidia GPUs, said Anshel Sag, principal analyst at Moor Strategy and Insights. “Many companies don’t necessarily want to send their workloads to the cloud. But also, they don’t want to build supercomputers to run models for inference,” Sag said. GenAI models are also becoming leaner and more efficient to run locally, making it possible to run reasoning models efficiently on workstations, Sag said. “DeepSeek was definitely one of the catalysts of that,” he said. DeepSeek was developed by the namesake Chinese company and has been widely adopted by US AI companies and cloud providers. During the Computex keynote, Nvidia CEO Huang acknowledged the role of DeepSeek in making AI accessible. “DeepSeek R1 is genuinely a gift to the world’s AI industry… R1 has made a real impact in how people think about AI, how to think about inference and how to think about reasoning,” Huang said.